Deep Reinforcement Learning

Transcription

1 Deep Reinforcement Learning Sanket Lokegaonkar Advanced Computer Vision (ECE 6554)

2 Outline The Why? Gliding Over All : An Introduction Classical RL DQN-Era Playing Atari with Deep Reinforcement Learning [2013] 7 Atari Games Human-level control through deep reinforcement learning. [2015] 49 Atari Games Brave New World

3 The Why? : Task Learning to behave optimally in a changing world Characteristics of the Task: RL: No Supervisor ( Only Rewards) Delayed Feedback Non I.I.D data Previous action affects the next state Learning by Interaction and your choices Time-travelling back to the moment: Only to Live Again Reminds you of something: Life - (Time Travelling Part)

4 The Why?: Hardness Imagine playing a new game whose rules you don t know; after a hundred hundred or so moves, your opponent announces, You lose. Russell and Norvig Introduction to Artificial Intelligence

5 The Why?: Why do RL? So that we could move away from rule-based systems!! So that we could train quadcopters to fly and move! So that you could sip coffee all day and your bot will do the trading for you! Or Create intelligent robots for dumb tasks

6 Gliding Over All: An Introduction Credits: David Silver

7 Environment and Agent At each step t, the agent The environment, Executes action At Receives observation Ot Receives scalar reward Rt Receives action At Emits observation Ot+1 Emits scalar reward Rt+1 t increments in every step Credits: David Silver

8 Markov Decision Process Credits: Andrej Karpathy s blog

9 Markov Decision Process (MDP) Set of states S Set of actions A State transition probabilities p(s s, a). This is the probability distribution over the state space given we take action a in state s Discount factor γ in [0, 1] Reward function R: S x A -> set of real numbers For simplicity, assume discrete rewards Finite MDP if both S and A are finite

10 State Transition Probabilities Suppose the reward function is discrete and maps from S x A to W The state transition probability or probability of transitioning to state s given current state s and action a in that state is given by:

11 Expected Rewards The expected reward for a given state-action pair is given by:

12 Policy and Value Function Policy Behavior Function Mapping from state to action Deterministic Policy : a=\pi (s) Stochastic Policy: (a s) = Pr [ At = a St =s] Value Function Prediction of total future Reward under a policy Captures goodness/badness of states Can be used to greedily select among the actions v (s) = E [R t+1 + Rt R t+3 St =s] Credits: ICML DRL Tutorial

13 Action-Value Function (Q -function) Expected Return starting from state s, taking action A and then following policy with as the discounting factor. Goodness of state given an action a Credits: ICML DRL Tutorial

14 Optimal Action-Value Function Action -Value Function Equivalent Representation (Bellman Equation): Optimal Value Function: Q* (s,a) = max Q (s,a) If we know the optimal action-value function, the optimal policy would be: Q (s,a) = E [ Rs + γ Q (s,a ) ] * ( s) = argmax a Q*(s,a) Optimal value maximizes over all decisions: Q* (s,a) = E [ Rs + γ maxa Q (s,a ) ] Credits: ICML DRL Tutorial

15 Bellman Equation (1) The equation expresses the relationship between the value of a state s and the values of its successor states The value of the next state must equal the discounted value of the expected next state, plus the reward expected along the way

16 Bellman Equation (2) The value of state s is the expected value of the sum of time-discounted rewards (starting at current state) given current state s This is expected value of r plus the sum of time-discounted rewards (starting at successor state) over all successor states s and all next rewards r and over all possible actions a in current state s

17 DEMO Credits: [3]

18 Approaches in RL Value -Based RL (Value Iteration) (Model Free) Policy -Based RL (Policy Iteration) (Model Free) Estimate the value of optimal value function Q* (s,a) Search for the optimal policy * ( s) Model-Based RL Learn a model of the environment Plan using the model Deep networks are used to approximate the functions

19 Model-free versus Model-based A model of the environment allows inferences to be made about how the environment will behave Example: Given a state and an action to be taken while in that state, the model could predict the next state and the next reward Models are used for planning, which means deciding on a course of action by considering possible future situations before they are experienced Model-based methods use models and planning. Think of this as modelling the dynamics p(s s, a) Model-free methods learn exclusively from trial-and-error (i.e. no modelling of the environment) We focus on model-free methods today:

20 On-policy versus Off-policy An on-policy agent learns only about the policy that it is executing An off-policy agent learns about a policy or policies different from the one that it is executing

21 What is TD learning? Temporal-Difference learning = TD learning The prediction problem is that of estimating the value function for a policy π The control problem is the problem of finding an optimal policy π* Given some experience following a policy π, update estimate v of vπ for non-terminal states occurring in that experience Given current step t, TD methods wait until the next time step to update V(St) Learn from partial returns

22 Epsilon-greedy Policy At each time step, the agent selects an action The agent follows the greedy strategy with probability 1 epsilon The agent selects a random action with probability epsilon With Q-learning, the greedy strategy is the action a that maximises Q given St+1

23 Q-learning Off-policy TD Control Similar to SARSA but off-policy updates The learned action-value function Q directly approximates the optimal action-value function q* independent of the policy being followed In update rule, choose action a that maximises Q given St+1 and use the resulting Q-value (i.e. estimated value given by optimal action-value function) plus the observed reward as the target This method is off-policy because we do not have a fixed policy that maps from states to actions. This is why At+1 is not used in the update rule

24 Deep Q-Networks (DQN) Introduced deep reinforcement learning It is common to use a function approximator Q(s, a; θ) to approximate the action-value function in Q-learning Deep Q-Networks is Q-learning with a deep neural network function approximator called the Q-network Discrete and finite set of actions A Example: Breakout has 3 actions move left, move right, no movement Uses epsilon-greedy policy to select actions

25 Q-Networks Core idea: We want the neural network to learn a non-linear hierarchy of features or feature representation that gives accurate Q-value estimates The neural network has a separate output unit for each possible action, which gives the Q-value estimate for that action given the input state The neural network is trained using mini-batch stochastic gradient updates and experience replay

26 State representation It is difficult to give the neural network a sequence of arbitrary length as input Use fixed length representation of sequence/history produced by a function ϕ(st) Example: The last 4 image frames in the sequence of Breakout gameplay

27 Q-Network Training Sample random mini-batch of experience tuples uniformly at random from D Similar to Q-learning update rule but: Use mini-batch stochastic gradient updates The gradient of the loss function for a given iteration with respect to the parameter θi is the difference between the target value and the actual value is multiplied by the gradient of the Q function approximator Q(s, a; θ) with respect to that specific parameter Use the gradient of the loss function to update the Q function approximator

28 Value-based RL Q-function is represented as Q-network with weights : Q(s,a,w) = Q*(s,a)

29 Playing Atari with Deep Reinforcement Learning Problem Statement: Input: Output: Raw Atari Frames 210 x160 pixel with 128-color pallete Possible Action to be taken Goal /Objective: Optimal Policy with Maximal Reward

30 Playing Atari with Deep Reinforcement Learning Architecture:

31 Playing Atari with Deep Reinforcement Learning Architecture:

32 Playing Atari with Deep Reinforcement Learning Architecture:

33 Playing Atari with Deep Reinforcement Learning Algorithm:

34 Playing Atari with Deep Reinforcement Learning MSE Loss with DQN: Issues with Convergence of DQN: Non I.I.D data Oscillating policies with slight variations of Q-values Gradient descent can be large and unstable

35 Solutions to the Issues: Non-IID data Experience Replay (for handling non-iid data) Build a memory storing N pairs of agent s experience i.e (st,at,rt+1,st+1) Sample random minibatch experience from the memory Breaks correlation ~ brings back to iid domain

36 Solution to the Issues: How to prevent oscillations Fix parameters used in Q-learning target Compute Q-learning targets wrt old fixed parameters θ Loss Minimization Equation between Q-network and Q-learning targets L(θ) = Es,a,r,s ~ D [ r + ᵞ maxa Q(s,a,θ) ] Periodically update fixed parameters θ-

37 Solution to the Issues: Unstable Gradient Clipping Rewards between [-1,1] Limits the scale of the derivatives Easier to use same learning rate over different Atari Games Disadvantage: Invariant to different magnitudes of rewards

38 Evaluations Atari Games Tried: Beam Rider Breakout Pong Q*bert Seaquest Space Invaders

39 Evaluations Metrics: Average Total Reward Metric Average action-value Collect a fixed set of states by running a random policy before training starts Average of the maximum predicted Q for these states Optimization: RMSProp with minibatches of size 32 Behavior policy ϵ greedy (Annealed from 1 to 0.1 and fixed to 0.1 later)

40 Evaluations

41 Evaluations Average Total Reward Best Performing Episode

42 Analysis Strengths: Q-learning over non-linear function approximators End-to-End Learning over multiple games SGD Training Seminal paper Weakness: DQN limited to finite discrete actions Long Training Reward Clipping

43 Human-level control through deep reinforcement learning Extending the DQN architecture to play 49 Atari 2600 arcade games No pretraining No game-specific training State: Transitions from 4 frames : Experience Replay Actions -18: 9 directions of joystick 9 directions + button Reward - Game Score

44 Architecture

45 Stats over 49 games

46 T-SNE Embedding ( Last Hidden Layer)

47 Analysis Strengths: End-to-End Learning over multiple games Beats human performance on most of the games Richer Rewards Weakness: Long Training DQN limited to finite discrete actions

48 Brave New World

49 Optimizations on DQN Since Then Double DQN: Remove Upward bias caused by maxa Q(s,a, w) Prioritized replay : Weight experience according to TD-error (suprise) Current Q-network w: Used to select actions Older Q-network w-: Used to evaluate actions Store experience in priority queue according to DQN error Asynchronous RL Joint Training using Parameter Sharing on Distributed Scale GORILA Credits: ICML DRL Tutorial

50 What are Policy Gradient Methods? Before: Learn the values of actions and then select actions based on their estimated action-values. The policy was generated directly from the value function We want to learn a parameterised policy that can select actions without consulting a value function. The parameters of the policy are called policy weights A value function may be used to learn the policy weights but this is not required for action selection Policy gradient methods are methods for learning the policy weights using the gradient of some performance measure with respect to the policy weights Policy gradient methods seek to maximise performance and so the policy weights are updated using gradient ascent

51 Policy-based Reinforcement Learning Search directly for the optimal policy π* Can use any parametric supervised machine learning model to learn policies π(a s; θ) where θ represents the learned parameters Recall that the optimal policy is the policy that achieves maximum future return

52 Policy-Based RL Represent policies by deep networks instead of Q-function Objective function for the network a = π( a s, u) Stochastic Policies A = π( s, u ) Deterministic Policies where u is the parameters of the deep network L(u) = E [ r 1 + γr 2 + γ2r 3 π(., u ) ] SGD Optimization Allows Continuous and Discrete Control Known to get stuck in Local minima Credits: ICML DRL Tutorial

53 Algorithms in Policy-Based RL REINFORCE while (true) run_episode(policy) update(policy) end; Actor-Critic Episodic updates Maximize the loss of expected reward under the objective Updates at each step Critic approximates the value function Actor approximates the policy Asynchronous Advantage Function Actor-Critic Uses Advantage Function Estimate for state-actuib. A(s,a,w) = Q(s,a,w) -V(s) Replacing the need of replay memory by using parallel agents running on CPU Relies on different exploration behavior of the parallel agents Outperforms the conventional method Credits: ICML DRL Tutorial

54 What is Asynchronous Reinforcement Learning? Use asynchronous gradient descent to optimise controllers This is useful for deep reinforcement learning where the controllers are deep neural networks, which take a long time to train Asynchronous gradient descent speeds up the learning process Can use one multi-core CPU to train deep neural networks asynchronously instead of multiple GPUs

55 Parallelism (1) Asynchronously execute multiple agents in parallel on multiple instances of the environment This parallelism decorrelates the agents data into a more stationary process since at any given time-step, the agents will be experiencing a variety of different states This approach enables a larger spectrum of fundamental on-policy and off-policy reinforcement learning algorithms to be applied robustly and effectively using deep neural networks Use asynchronous actor-learners (i.e. agents). Think of each actor-learner as a thread Run everything on a single multi-core CPU to avoid communication costs of sending gradients and parameters

56 Parallelism (2) Multiple actor-learners running in parallel are likely to be exploring different parts of the environment We can explicitly use different exploration policies in each actor-learner to maximise this diversity By running different exploration policies in different threads, the overall changes made to the parameters by multiple actor-learners applying updates in parallel are less likely to be correlated in time than a single agent applying online updates

57 No Experience Replay No need for a replay memory. We instead rely on parallel actors employing different exploration policies to perform the stabilising role undertaken by experience replay in the DQN training algorithm Since we no longer rely on experience replay for stabilising learning, we are able to use on-policy reinforcement learning methods to train neural networks in a stable way

58 Video Demo : A3C Labryinth

59 Video Demo : DQN Doom

60 Scope / Future Multi-agent Deep RL Share Parameters!!! (Naive approach) Hierarchical Deep Reinforcement Learning Road to General AI!

61 Quotes from the Maestro If intelligence was a cake, unsupervised learning would be the cake, supervised learning would be the icing on the cake, and reinforcement learning would be the cherry on the cake. -Yann Lecun

62 Reference Zone ICML Deep RL Tutorial : Andrej Karpathy s blog: Playing Atari with Deep Reinforcement Learning. V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, M. Riedmiller, NIPS workshop 2013 Human-level control through deep reinforcement learning. Mnih et al. Nature